Github user squito commented on the issue:
https://github.com/apache/spark/pull/11105
hi @holdenk sorry for such a ridiculously long delay from me. I'll just
make some high-level comments first, and make another pass with more particular
comments on the code.
After reminding myself about all the issues & tradeoffs that you needed to
make, I think this is looking really good. My only real concern is the extra
magic that needs to happen in `MapPartitionsRDD`. What happens with a custom
user RDD, which doesn't have that magic in it? I'm guessing data property
accumulators will just behave like regular accumulators? It would be nice if
we could warn the user about that, but maybe its not possible.
General things that I think need to be done:
* internal docs & code clarity (I will make more detailed comments)
* I feel like there are probably some more test cases needed, with
interleaved execution of shared rdds etc. I'll think about it and make some
suggestions here, its possible the existing tests are good I just need to walk
through them a bit more carefully.
* `@Experimental` on everything using this api. I know you've mentioned it
in the api docs, but I think you should probably explicitly label all methods
which expose it at all (may require changing some default-arg methods to be
overloads instead).
* user-facing docs. there is a little blurb there but I think we need a
better explanation of when to use each one. (you can hold off on this a bit.)
* tests involving datasets. I don't expect any extra issues here but I
think its worth having them in place.
* some measure of the performance effect of using these. Its great you've
found a solution which avoids any overhead when data-property accumulators
aren't used. But it would be good to have a *rough* idea of what kind of
overhead we're talking about when you do use them -- if its abyssmally slow
we'll have to rethink it. (I really don't expect that to be the case, though.)
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
